A multi-source feature-driven deep learning method to generate linear energy transfer distribution for proton...

Area of Science:

• Medical Physics
• Radiotherapy
• Computational Biology

Background:

• Monte Carlo (MC) simulations provide accurate dose-averaged linear energy transfer (LETd) estimation in proton therapy but are computationally intensive.
• Deep learning (DL) models offer faster alternatives, yet few are optimized for multi-source inputs or validated on complex tumors like nasopharyngeal carcinoma (NPC).

Purpose of the Study:

• Develop and evaluate a multi-source DL framework for voxel-level LETd prediction in proton therapy for NPC.
• Identify optimal input configurations for DL models and assess prediction accuracy and clinical reliability.

Main Methods:

• Analyzed 100 NPC patients (500 proton therapy fields) with ground-truth LETd maps from GPU-based MC simulations.
• Trained a DL model with eight configurations using dose distribution, CT images, beam masks, and radiological depth.
• Evaluated performance using MAE, PSNR, NCC, SSIM, and 3D Gamma analysis; compared with a 3D U-Net baseline.

Main Results:

• The M_dose+ct DL model achieved superior performance (MAE 0.21 ± 0.03 keV/µm) compared to the 3D U-Net baseline (MAE 0.38 ± 0.52 keV/µm).
• Model interpretability (Grad-CAM) confirmed focus on relevant regions, and uncertainty maps identified high-variability zones.
• DL model predictions were generated in ~7 seconds per case, significantly faster than MC simulations (~30 seconds).

Conclusions:

• The multi-source DL framework enables rapid and accurate voxel-level LETd prediction for NPC proton therapy.
• Combining dose and CT data proved to be the most reliable input configuration.
• Interpretability and uncertainty analyses enhance the framework's suitability for biologically guided proton therapy planning.